m3taversal
e830fe4c5f
Three-agent knowledge base (Leo, Rio, Clay) with: - 177 claim files across core/ and foundations/ - 38 domain claims in internet-finance/ - 22 domain claims in entertainment/ - Agent soul documents (identity, beliefs, reasoning, skills) - 14 positions across 3 agents - Claim/belief/position schemas - 6 shared skills - Agent-facing CLAUDE.md operating manual Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
76 lines
2.7 KiB
Markdown
76 lines
2.7 KiB
Markdown
# Skill: Evaluate
|
|
|
|
Multi-agent evaluation of proposed claims before they enter the shared knowledge base.
|
|
|
|
## When to Use
|
|
|
|
When candidate claims exist in inbox/proposals/ awaiting review.
|
|
|
|
## Process
|
|
|
|
### Step 1: Leo assigns evaluation
|
|
|
|
Leo reviews the proposed claim and identifies:
|
|
- Primary domain (which agent is the lead evaluator?)
|
|
- Secondary domains (which other agents should weigh in?)
|
|
- Urgency (time-sensitive? can wait for full review cycle?)
|
|
|
|
### Step 2: Domain agents evaluate
|
|
|
|
Each assigned agent reviews the claim against these criteria:
|
|
|
|
**Quality checks:**
|
|
1. Is this specific enough to disagree with?
|
|
2. Is the evidence traceable and verifiable?
|
|
3. Does the description add information beyond the title?
|
|
4. Is the confidence level appropriate for the evidence strength?
|
|
|
|
**Knowledge base checks:**
|
|
5. Does this duplicate an existing claim? (cite the existing one if so)
|
|
6. Does it contradict an existing claim? (if so, is the contradiction explicit and argued?)
|
|
7. Does it add genuine value the knowledge base doesn't already have?
|
|
8. Are wiki links pointing to real files?
|
|
|
|
**Domain-specific evaluation:**
|
|
9. Does this match the agent's understanding of the domain landscape?
|
|
10. Would this change any of the agent's current beliefs or positions?
|
|
11. Are there cross-domain implications other agents should know about?
|
|
|
|
### Step 3: Agents vote
|
|
|
|
Each evaluating agent submits one of:
|
|
- **Accept** — claim meets all quality criteria, add to knowledge base
|
|
- **Accept with changes** — good claim but needs specific modifications (list them)
|
|
- **Reject** — fails quality criteria (explain which ones and why)
|
|
- **Request more evidence** — interesting claim but insufficient evidence to accept
|
|
|
|
### Step 4: Leo synthesizes
|
|
|
|
- If consensus accept: merge into knowledge base
|
|
- If consensus reject: close with explanation
|
|
- If mixed: Leo synthesizes the disagreement
|
|
- Factual disagreement → identify what evidence would resolve it
|
|
- Perspective disagreement → both interpretations may be valid
|
|
- Quality concerns → specific changes needed
|
|
- If request more evidence: assign research task to relevant agent
|
|
|
|
### Step 5: Post-merge cascade check
|
|
|
|
After a claim is accepted:
|
|
- Does this affect any agent's beliefs? (check depends_on chains)
|
|
- Flag affected beliefs as needs_review
|
|
- Notify owning agents
|
|
|
|
## Output
|
|
|
|
- Evaluation record: which agents reviewed, how they voted, outcome
|
|
- Merged claim (if accepted) in domains/{domain}/
|
|
- Cascade flags (if applicable)
|
|
- Research tasks (if more evidence needed)
|
|
|
|
## Quality Gate
|
|
|
|
- Every rejection explains which criteria failed
|
|
- Every mixed vote gets Leo synthesis
|
|
- Cascade checks run on every accepted claim
|
|
- Evaluation record is preserved for transparency
|